chunk_id
int64
0
448
chunk_text
stringlengths
1
10.8k
chunk_text_tokens
int64
1
2.01k
serialized_text
stringlengths
1
11.1k
serialized_text_tokens
int64
1
2.02k
31
(stochastic optimization). For probabilities p=(p1,…,pm)∈P𝑝subscript𝑝1…subscript𝑝𝑚𝑃p=(p_{1},\dots,p_{m})\in P, where P⊂ℝm𝑃superscriptℝ𝑚P\subset\mathbb{R}^{m} is the set of nonnegative vectors with components summing to one, consider the problem , 1 = minimizex∈X∑i=1mpi​fi​(x),subscriptminimize𝑥𝑋superscriptsubs...
916
3 Examples 3.2 Example (stochastic optimization). For probabilities p=(p1,…,pm)∈P𝑝subscript𝑝1…subscript𝑝𝑚𝑃p=(p_{1},\dots,p_{m})\in P, where P⊂ℝm𝑃superscriptℝ𝑚P\subset\mathbb{R}^{m} is the set of nonnegative vectors with components summing to one, consider the problem , 1 = minimizex∈X∑i=1mpi​fi​(x),subscriptmini...
924
32
(distributionally robust optimization). Let P⊂ℝm𝑃superscriptℝ𝑚P\subset\mathbb{R}^{m} be as in the previous example and let A𝐴A and Aνsuperscript𝐴𝜈A^{\nu} be nonempty closed subsets of P𝑃P. We consider the problem , 1 = minimizex∈Xmaxp∈A∑i=1mpi​fi​(x)subscriptminimize𝑥𝑋subscriptmax𝑝𝐴superscriptsubscript𝑖1𝑚su...
1,784
3 Examples 3.3 Example (distributionally robust optimization). Let P⊂ℝm𝑃superscriptℝ𝑚P\subset\mathbb{R}^{m} be as in the previous example and let A𝐴A and Aνsuperscript𝐴𝜈A^{\nu} be nonempty closed subsets of P𝑃P. We consider the problem , 1 = minimizex∈Xmaxp∈A∑i=1mpi​fi​(x)subscriptminimize𝑥𝑋subscriptmax𝑝𝐴supe...
1,792
33
Since there are p¯ν∈Asuperscript¯𝑝𝜈𝐴\bar{p}^{\nu}\in A such that ‖pν−p¯ν‖2→0→subscriptnormsuperscript𝑝𝜈superscript¯𝑝𝜈20\|p^{\nu}-\bar{p}^{\nu}\|_{2}\to 0 because Aν​→s ​Asuperscript𝐴𝜈→s 𝐴A^{\nu}\,{\lower 1.0pt\hbox{$\rightarrow$}}\kern-10.0pt\hbox{\raise 4.0pt\hbox{$\,\scriptstyle s$}}\hskip 7.0ptA, limsuphν​...
336
3 Examples 3.3 Example Since there are p¯ν∈Asuperscript¯𝑝𝜈𝐴\bar{p}^{\nu}\in A such that ‖pν−p¯ν‖2→0→subscriptnormsuperscript𝑝𝜈superscript¯𝑝𝜈20\|p^{\nu}-\bar{p}^{\nu}\|_{2}\to 0 because Aν​→s ​Asuperscript𝐴𝜈→s 𝐴A^{\nu}\,{\lower 1.0pt\hbox{$\rightarrow$}}\kern-10.0pt\hbox{\raise 4.0pt\hbox{$\,\scriptstyle s$}}\...
344
34
(augmented Lagrangian methods). With h​(z)=z1+∑i=2mι{0}​(zi)ℎ𝑧subscript𝑧1superscriptsubscript𝑖2𝑚subscript𝜄0subscript𝑧𝑖h(z)=z_{1}+\sum_{i=2}^{m}\iota_{\{0\}}(z_{i}), the problem , 1 = minimizex∈Xf1​(x)​ subject to ​f2​(x)=0,…,fm​(x)=0formulae-sequencesubscriptminimize𝑥𝑋subscript𝑓1𝑥 subject to subscript𝑓2𝑥0…...
1,412
3 Examples 3.4 Example (augmented Lagrangian methods). With h​(z)=z1+∑i=2mι{0}​(zi)ℎ𝑧subscript𝑧1superscriptsubscript𝑖2𝑚subscript𝜄0subscript𝑧𝑖h(z)=z_{1}+\sum_{i=2}^{m}\iota_{\{0\}}(z_{i}), the problem , 1 = minimizex∈Xf1​(x)​ subject to ​f2​(x)=0,…,fm​(x)=0formulae-sequencesubscriptminimize𝑥𝑋subscript𝑓1𝑥 subj...
1,420
35
(min-functions). For smooth gi​k:ℝn→ℝ:subscript𝑔𝑖𝑘→superscriptℝ𝑛ℝg_{ik}:\mathbb{R}^{n}\to\mathbb{R}, k=1,…,si𝑘1…subscript𝑠𝑖k=1,\dots,s_{i}, i=1,…,m𝑖1…𝑚i=1,\dots,m, let , 1 = fi​(x)=mink=1,…,si⁡gi​k​(x).subscript𝑓𝑖𝑥subscript𝑘1…subscript𝑠𝑖subscript𝑔𝑖𝑘𝑥f_{i}(x)=\min_{k=1,\dots,s_{i}}g_{ik}(x).. , 2 = W...
1,673
3 Examples 3.5 Example (min-functions). For smooth gi​k:ℝn→ℝ:subscript𝑔𝑖𝑘→superscriptℝ𝑛ℝg_{ik}:\mathbb{R}^{n}\to\mathbb{R}, k=1,…,si𝑘1…subscript𝑠𝑖k=1,\dots,s_{i}, i=1,…,m𝑖1…𝑚i=1,\dots,m, let , 1 = fi​(x)=mink=1,…,si⁡gi​k​(x).subscript𝑓𝑖𝑥subscript𝑘1…subscript𝑠𝑖subscript𝑔𝑖𝑘𝑥f_{i}(x)=\min_{k=1,\dots,s_{...
1,681
36
Since μi​kν​(x)∈(0,1)subscriptsuperscript𝜇𝜈𝑖𝑘𝑥01\mu^{\nu}_{ik}(x)\in(0,1) regardless of x𝑥x, we have that {∇fiν​(xν),ν∈ℕ}∇superscriptsubscript𝑓𝑖𝜈superscript𝑥𝜈𝜈ℕ\{\nabla f_{i}^{\nu}(x^{\nu}),\nu\in\mathbb{N}\} is bounded. For some N∈𝒩∞#𝑁subscriptsuperscript𝒩#N\in{\cal N}^{\scriptscriptstyle\#}_{\infty}, s...
1,904
3 Examples 3.5 Example Since μi​kν​(x)∈(0,1)subscriptsuperscript𝜇𝜈𝑖𝑘𝑥01\mu^{\nu}_{ik}(x)\in(0,1) regardless of x𝑥x, we have that {∇fiν​(xν),ν∈ℕ}∇superscriptsubscript𝑓𝑖𝜈superscript𝑥𝜈𝜈ℕ\{\nabla f_{i}^{\nu}(x^{\nu}),\nu\in\mathbb{N}\} is bounded. For some N∈𝒩∞#𝑁subscriptsuperscript𝒩#N\in{\cal N}^{\scriptscr...
1,912
37
, 1 = ∇fiν​(xν)=∑k=1siμi​kν​(xν)​∇gi​k​(xν)​→N ​v¯=∑k∈𝔸i​(x¯)μi​k∞​∇gi​k​(x¯)∈con​∂fi​(x¯)∇subscriptsuperscript𝑓𝜈𝑖superscript𝑥𝜈superscriptsubscript𝑘1subscript𝑠𝑖subscriptsuperscript𝜇𝜈𝑖𝑘superscript𝑥𝜈∇subscript𝑔𝑖𝑘superscript𝑥𝜈→N ¯𝑣subscript𝑘subscript𝔸𝑖¯𝑥subscriptsuperscript𝜇𝑖𝑘∇subscript𝑔𝑖𝑘¯�...
481
3 Examples 3.5 Example , 1 = ∇fiν​(xν)=∑k=1siμi​kν​(xν)​∇gi​k​(xν)​→N ​v¯=∑k∈𝔸i​(x¯)μi​k∞​∇gi​k​(x¯)∈con​∂fi​(x¯)∇subscriptsuperscript𝑓𝜈𝑖superscript𝑥𝜈superscriptsubscript𝑘1subscript𝑠𝑖subscriptsuperscript𝜇𝜈𝑖𝑘superscript𝑥𝜈∇subscript𝑔𝑖𝑘superscript𝑥𝜈→N ¯𝑣subscript𝑘subscript𝔸𝑖¯𝑥subscriptsuperscript�...
489
38
(penalty methods). With h​(z)=z1+ι(−∞,0]m−1​(z2,…,zm)ℎ𝑧subscript𝑧1subscript𝜄superscript0𝑚1subscript𝑧2…subscript𝑧𝑚h(z)=z_{1}+\iota_{(-\infty,0]^{m-1}}(z_{2},\dots,z_{m}), we consider the problem , 1 = minimizex∈Xf1​(x)​ subject to ​f2​(x)≤0,…,fm​(x)≤0,formulae-sequencesubscriptminimize𝑥𝑋subscript𝑓1𝑥 subject t...
1,254
3 Examples 3.6 Example (penalty methods). With h​(z)=z1+ι(−∞,0]m−1​(z2,…,zm)ℎ𝑧subscript𝑧1subscript𝜄superscript0𝑚1subscript𝑧2…subscript𝑧𝑚h(z)=z_{1}+\iota_{(-\infty,0]^{m-1}}(z_{2},\dots,z_{m}), we consider the problem , 1 = minimizex∈Xf1​(x)​ subject to ​f2​(x)≤0,…,fm​(x)≤0,formulae-sequencesubscriptminimize𝑥𝑋s...
1,262
39
(interior-point methods). For the actual problem in Example 3.6, we consider the approximation , 1 = hν​(z)={z1−1θν​∑i=2mln⁡(−zi) if ​z2<0,…,zm<0∞ otherwise,superscriptℎ𝜈𝑧casessubscript𝑧11superscript𝜃𝜈superscriptsubscript𝑖2𝑚subscript𝑧𝑖formulae-sequence if subscript𝑧20…subscript𝑧𝑚0 otherwise,h^{\nu}(z)=\begi...
1,161
3 Examples 3.7 Example (interior-point methods). For the actual problem in Example 3.6, we consider the approximation , 1 = hν​(z)={z1−1θν​∑i=2mln⁡(−zi) if ​z2<0,…,zm<0∞ otherwise,superscriptℎ𝜈𝑧casessubscript𝑧11superscript𝜃𝜈superscriptsubscript𝑖2𝑚subscript𝑧𝑖formulae-sequence if subscript𝑧20…subscript𝑧𝑚0 oth...
1,169
40
(expectation functions). Expectation functions arise in stochastic optimization and machine learning and then the component functions f1,…,fmsubscript𝑓1…subscript𝑓𝑚f_{1},\dots,f_{m} may take the form , 1 = fi​(x)=𝔼​[gi​(𝝃,x)],subscript𝑓𝑖𝑥𝔼delimited-[]subscript𝑔𝑖𝝃𝑥f_{i}(x)=\mathbb{E}\big{[}g_{i}(\mbox{\bold...
923
3 Examples 3.8 Example (expectation functions). Expectation functions arise in stochastic optimization and machine learning and then the component functions f1,…,fmsubscript𝑓1…subscript𝑓𝑚f_{1},\dots,f_{m} may take the form , 1 = fi​(x)=𝔼​[gi​(𝝃,x)],subscript𝑓𝑖𝑥𝔼delimited-[]subscript𝑔𝑖𝝃𝑥f_{i}(x)=\mathbb{E}\...
931
41
(oracle functions). Suppose that hℎh is real-valued, but not available in an explicit form. If for each z𝑧z we can compute h​(z)ℎ𝑧h(z) and a subgradient v∈∂h​(z)𝑣ℎ𝑧v\in\partial h(z), then the approximation , 1 = hν​(z)=maxk=1,…,ν⁡h​(zk)+⟨vk,z−zk⟩, with ​zk∈ℝm,vk∈∂h​(zk),formulae-sequencesuperscriptℎ𝜈𝑧subscript𝑘1...
966
3 Examples 3.9 Example (oracle functions). Suppose that hℎh is real-valued, but not available in an explicit form. If for each z𝑧z we can compute h​(z)ℎ𝑧h(z) and a subgradient v∈∂h​(z)𝑣ℎ𝑧v\in\partial h(z), then the approximation , 1 = hν​(z)=maxk=1,…,ν⁡h​(zk)+⟨vk,z−zk⟩, with ​zk∈ℝm,vk∈∂h​(zk),formulae-sequencesuper...
974
42
(homotopy method). For proper, lsc, and convex h^:ℝm−1→ℝ¯:^ℎ→superscriptℝ𝑚1¯ℝ\hat{h}:\mathbb{R}^{m-1}\to\overline{\mathbb{R}} and lLc F^:ℝn→ℝm−1:^𝐹→superscriptℝ𝑛superscriptℝ𝑚1\hat{F}:\mathbb{R}^{n}\to\mathbb{R}^{m-1}, consider the problem , 1 = minimizex∈Xh^​(F^​(x)),subscriptminimize𝑥𝑋^ℎ^𝐹𝑥\mathop{\rm minimize...
1,563
3 Examples 3.10 Example (homotopy method). For proper, lsc, and convex h^:ℝm−1→ℝ¯:^ℎ→superscriptℝ𝑚1¯ℝ\hat{h}:\mathbb{R}^{m-1}\to\overline{\mathbb{R}} and lLc F^:ℝn→ℝm−1:^𝐹→superscriptℝ𝑛superscriptℝ𝑚1\hat{F}:\mathbb{R}^{n}\to\mathbb{R}^{m-1}, consider the problem , 1 = minimizex∈Xh^​(F^​(x)),subscriptminimize𝑥𝑋^ℎ^...
1,571
43
(monitoring functions). In extended nonlinear programming [49], one utilizes , 1 = h​(z)=supy∈Y{⟨z,y⟩−12​⟨y,B​y⟩}​ and ​hν​(z)=supy∈Yν{⟨z,y⟩−12​⟨y,Bν​y⟩},ℎ𝑧subscriptsup𝑦𝑌𝑧𝑦12𝑦𝐵𝑦 and superscriptℎ𝜈𝑧subscriptsup𝑦superscript𝑌𝜈𝑧𝑦12𝑦superscript𝐵𝜈𝑦h(z)=\mathop{\rm sup}\nolimits_{y\in Y}\big{\{}\langle z,y\r...
1,465
3 Examples 3.11 Example (monitoring functions). In extended nonlinear programming [49], one utilizes , 1 = h​(z)=supy∈Y{⟨z,y⟩−12​⟨y,B​y⟩}​ and ​hν​(z)=supy∈Yν{⟨z,y⟩−12​⟨y,Bν​y⟩},ℎ𝑧subscriptsup𝑦𝑌𝑧𝑦12𝑦𝐵𝑦 and superscriptℎ𝜈𝑧subscriptsup𝑦superscript𝑌𝜈𝑧𝑦12𝑦superscript𝐵𝜈𝑦h(z)=\mathop{\rm sup}\nolimits_{y\in...
1,473
44
(difference-of-convex functions). For a proper, lsc, and convex f:ℝn→ℝ¯:𝑓→superscriptℝ𝑛¯ℝf:\mathbb{R}^{n}\to\overline{\mathbb{R}} and convex g:ℝn→ℝ:𝑔→superscriptℝ𝑛ℝg:\mathbb{R}^{n}\to\mathbb{R}, consider the problem , 1 = minimizex∈Xf​(x)−g​(x),subscriptminimize𝑥𝑋𝑓𝑥𝑔𝑥\mathop{\rm minimize}_{x\in X}f(x)-g(x),. ...
1,470
3 Examples 3.12 Example (difference-of-convex functions). For a proper, lsc, and convex f:ℝn→ℝ¯:𝑓→superscriptℝ𝑛¯ℝf:\mathbb{R}^{n}\to\overline{\mathbb{R}} and convex g:ℝn→ℝ:𝑔→superscriptℝ𝑛ℝg:\mathbb{R}^{n}\to\mathbb{R}, consider the problem , 1 = minimizex∈Xf​(x)−g​(x),subscriptminimize𝑥𝑋𝑓𝑥𝑔𝑥\mathop{\rm minimi...
1,478
45
As an example of an implementable version of the consistent approximation algorithm, we consider the setting where Xνsuperscript𝑋𝜈X^{\nu} is convex, hνsuperscriptℎ𝜈h^{\nu} is real-valued, and Fνsuperscript𝐹𝜈F^{\nu} is twice continuously differentiable. Then, the approximating problems (2.1) are solvable by proxima...
1,925
4 Enhanced Proximal Composite Algorithm As an example of an implementable version of the consistent approximation algorithm, we consider the setting where Xνsuperscript𝑋𝜈X^{\nu} is convex, hνsuperscriptℎ𝜈h^{\nu} is real-valued, and Fνsuperscript𝐹𝜈F^{\nu} is twice continuously differentiable. Then, the approximatin...
1,933
46
, 1 = y¯k+1∈∂hν​(z¯k+1)​ and −∇Fν​(x¯k)⊤​y¯k+1−1λk​(x¯k+1−x¯k)∈NXν​(x¯k+1).superscript¯𝑦𝑘1superscriptℎ𝜈superscript¯𝑧𝑘1 and ∇superscript𝐹𝜈superscriptsuperscript¯𝑥𝑘topsuperscript¯𝑦𝑘11superscript𝜆𝑘superscript¯𝑥𝑘1superscript¯𝑥𝑘subscript𝑁superscript𝑋𝜈superscript¯𝑥𝑘1\bar{y}^{k+1}\in\partial h^{\nu}(\bar...
957
4 Enhanced Proximal Composite Algorithm , 1 = y¯k+1∈∂hν​(z¯k+1)​ and −∇Fν​(x¯k)⊤​y¯k+1−1λk​(x¯k+1−x¯k)∈NXν​(x¯k+1).superscript¯𝑦𝑘1superscriptℎ𝜈superscript¯𝑧𝑘1 and ∇superscript𝐹𝜈superscriptsuperscript¯𝑥𝑘topsuperscript¯𝑦𝑘11superscript𝜆𝑘superscript¯𝑥𝑘1superscript¯𝑥𝑘subscript𝑁superscript𝑋𝜈superscript¯𝑥...
965
47
(enhanced proximal composite algorithm). Suppose that Xν​→s ​Xsuperscript𝑋𝜈→s 𝑋X^{\nu}\,{\lower 1.0pt\hbox{$\rightarrow$}}\kern-10.0pt\hbox{\raise 4.0pt\hbox{$\,\scriptstyle s$}}\hskip 7.0ptX with these sets being convex, hν​→e ​hsuperscriptℎ𝜈→e ℎh^{\nu}\,{\lower 1.0pt\hbox{$\rightarrow$}}\kern-10.0pt\hbox{\raise 4...
1,642
4 Enhanced Proximal Composite Algorithm 4.1 Theorem (enhanced proximal composite algorithm). Suppose that Xν​→s ​Xsuperscript𝑋𝜈→s 𝑋X^{\nu}\,{\lower 1.0pt\hbox{$\rightarrow$}}\kern-10.0pt\hbox{\raise 4.0pt\hbox{$\,\scriptstyle s$}}\hskip 7.0ptX with these sets being convex, hν​→e ​hsuperscriptℎ𝜈→e ℎh^{\nu}\,{\lower ...
1,656
48
, 1 = −∇Fν​(x¯k)⊤​y−1λk​(x⋆−x¯k)∈NXν​(x⋆)​ for some ​y∈∂hν​(Fν​(x¯k)+∇Fν​(x¯k)​(x⋆−x¯k)).∇superscript𝐹𝜈superscriptsuperscript¯𝑥𝑘top𝑦1superscript𝜆𝑘superscript𝑥⋆superscript¯𝑥𝑘subscript𝑁superscript𝑋𝜈superscript𝑥⋆ for some 𝑦superscriptℎ𝜈superscript𝐹𝜈superscript¯𝑥𝑘∇superscript𝐹𝜈superscript¯𝑥𝑘superscr...
1,018
4 Enhanced Proximal Composite Algorithm 4.1 Theorem , 1 = −∇Fν​(x¯k)⊤​y−1λk​(x⋆−x¯k)∈NXν​(x⋆)​ for some ​y∈∂hν​(Fν​(x¯k)+∇Fν​(x¯k)​(x⋆−x¯k)).∇superscript𝐹𝜈superscriptsuperscript¯𝑥𝑘top𝑦1superscript𝜆𝑘superscript𝑥⋆superscript¯𝑥𝑘subscript𝑁superscript𝑋𝜈superscript𝑥⋆ for some 𝑦superscriptℎ𝜈superscript𝐹𝜈supe...
1,032
49
For the sake of contradiction, suppose that the algorithm iterates indefinitely without generating the next (xν,yν,zν)superscript𝑥𝜈superscript𝑦𝜈superscript𝑧𝜈(x^{\nu},y^{\nu},z^{\nu}). Let {x¯k,k∈ℕ}superscript¯𝑥𝑘𝑘ℕ\{\bar{x}^{k},k\in\mathbb{N}\} be the resulting sequence, which has a cluster point in view of (4....
1,064
4 Enhanced Proximal Composite Algorithm 4.1 Theorem For the sake of contradiction, suppose that the algorithm iterates indefinitely without generating the next (xν,yν,zν)superscript𝑥𝜈superscript𝑦𝜈superscript𝑧𝜈(x^{\nu},y^{\nu},z^{\nu}). Let {x¯k,k∈ℕ}superscript¯𝑥𝑘𝑘ℕ\{\bar{x}^{k},k\in\mathbb{N}\} be the resultin...
1,078
50
An inverse problem in machine learning is that of determining an input to a collection of neural networks such that their outputs best match a given quantity [28, 30]. Specifically, we are given s𝑠s neural networks represented by the mappings Fi:ℝn0→ℝnq:subscript𝐹𝑖→superscriptℝsubscript𝑛0superscriptℝsubscript𝑛𝑞F_...
1,716
5 Inverse Problems in Machine Learning An inverse problem in machine learning is that of determining an input to a collection of neural networks such that their outputs best match a given quantity [28, 30]. Specifically, we are given s𝑠s neural networks represented by the mappings Fi:ℝn0→ℝnq:subscript𝐹𝑖→superscriptℝ...
1,724
51
It is convenient to express (5.1) using additional variables as follows: We think of x1,i,k∈ℝnksubscript𝑥1𝑖𝑘superscriptℝsubscript𝑛𝑘x_{1,i,k}\in\mathbb{R}^{n_{k}} as the output of the k𝑘kth layer for neural network i𝑖i. Let r=∑k=1qnk𝑟superscriptsubscript𝑘1𝑞subscript𝑛𝑘r=\sum_{k=1}^{q}n_{k}, X=X^×ℝs​r𝑋^𝑋supe...
1,798
5 Inverse Problems in Machine Learning It is convenient to express (5.1) using additional variables as follows: We think of x1,i,k∈ℝnksubscript𝑥1𝑖𝑘superscriptℝsubscript𝑛𝑘x_{1,i,k}\in\mathbb{R}^{n_{k}} as the output of the k𝑘kth layer for neural network i𝑖i. Let r=∑k=1qnk𝑟superscriptsubscript𝑘1𝑞subscript𝑛𝑘r=...
1,806
52
(weak consistency in inverse machine learning). In the notation of this section, suppose that for each (i,j,k)𝑖𝑗𝑘(i,j,k), the following property holds: , 1 = γν∈ℝ→γαν∈∂gi,j,kν​(γν)}⟹{gi,j,kν​(γν)→gi,j,k​(γ){αν,ν∈ℕ}​ is bounded with all its cluster points in ​∂gi,j,k​(γ).⟹casessuperscript𝛾𝜈ℝ→𝛾otherwisesuperscript�...
1,872
5 Inverse Problems in Machine Learning 5.1 Proposition (weak consistency in inverse machine learning). In the notation of this section, suppose that for each (i,j,k)𝑖𝑗𝑘(i,j,k), the following property holds: , 1 = γν∈ℝ→γαν∈∂gi,j,kν​(γν)}⟹{gi,j,kν​(γν)→gi,j,k​(γ){αν,ν∈ℕ}​ is bounded with all its cluster points in ​∂gi...
1,885
53
Thus, we have Hiν​(xν)=0superscriptsubscript𝐻𝑖𝜈superscript𝑥𝜈0H_{i}^{\nu}(x^{\nu})=0 and xν→x¯→superscript𝑥𝜈¯𝑥x^{\nu}\to\bar{x}. Since h^^ℎ\hat{h} is real-valued and convex, h^ν​(x1,1,qν,…,x1,s,qν)superscript^ℎ𝜈superscriptsubscript𝑥11𝑞𝜈…superscriptsubscript𝑥1𝑠𝑞𝜈\hat{h}^{\nu}(x_{1,1,q}^{\nu},\dots,x_{1,s,...
1,626
5 Inverse Problems in Machine Learning 5.1 Proposition Thus, we have Hiν​(xν)=0superscriptsubscript𝐻𝑖𝜈superscript𝑥𝜈0H_{i}^{\nu}(x^{\nu})=0 and xν→x¯→superscript𝑥𝜈¯𝑥x^{\nu}\to\bar{x}. Since h^^ℎ\hat{h} is real-valued and convex, h^ν​(x1,1,qν,…,x1,s,qν)superscript^ℎ𝜈superscriptsubscript𝑥11𝑞𝜈…superscriptsubscr...
1,639
54
Since uν=Fν​(xν)−zνsuperscript𝑢𝜈superscript𝐹𝜈superscript𝑥𝜈superscript𝑧𝜈u^{\nu}=F^{\nu}(x^{\nu})-z^{\nu}, we conclude that u¯=F​(x¯)−z¯¯𝑢𝐹¯𝑥¯𝑧\bar{u}=F(\bar{x})-\bar{z}; recall that gi,j,kν​(γν)→gi,j,k​(γ)→superscriptsubscript𝑔𝑖𝑗𝑘𝜈superscript𝛾𝜈subscript𝑔𝑖𝑗𝑘𝛾g_{i,j,k}^{\nu}(\gamma^{\nu})\to g_{i,j...
1,670
5 Inverse Problems in Machine Learning 5.1 Proposition Since uν=Fν​(xν)−zνsuperscript𝑢𝜈superscript𝐹𝜈superscript𝑥𝜈superscript𝑧𝜈u^{\nu}=F^{\nu}(x^{\nu})-z^{\nu}, we conclude that u¯=F​(x¯)−z¯¯𝑢𝐹¯𝑥¯𝑧\bar{u}=F(\bar{x})-\bar{z}; recall that gi,j,kν​(γν)→gi,j,k​(γ)→superscriptsubscript𝑔𝑖𝑗𝑘𝜈superscript𝛾𝜈sub...
1,683
55
where ci,j,kν=∇ai,j,kν​(x)subscriptsuperscript𝑐𝜈𝑖𝑗𝑘∇subscriptsuperscript𝑎𝜈𝑖𝑗𝑘𝑥c^{\nu}_{i,j,k}=\nabla a^{\nu}_{i,j,k}(x). The assumption (5.2) implies that any sequence {di,j,kν∈∂fi,j,kν​(xν),ν∈N}formulae-sequencesuperscriptsubscript𝑑𝑖𝑗𝑘𝜈subscriptsuperscript𝑓𝜈𝑖𝑗𝑘superscript𝑥𝜈𝜈𝑁\{d_{i,j,k}^{\nu}\...
584
5 Inverse Problems in Machine Learning 5.1 Proposition where ci,j,kν=∇ai,j,kν​(x)subscriptsuperscript𝑐𝜈𝑖𝑗𝑘∇subscriptsuperscript𝑎𝜈𝑖𝑗𝑘𝑥c^{\nu}_{i,j,k}=\nabla a^{\nu}_{i,j,k}(x). The assumption (5.2) implies that any sequence {di,j,kν∈∂fi,j,kν​(xν),ν∈N}formulae-sequencesuperscriptsubscript𝑑𝑖𝑗𝑘𝜈subscriptsup...
597
56
Consistency furnishes guarantees about the limiting behavior of approximations, but it also can be beneficial to quantify the rate of convergence. In this section, we refine results from [53] and estimate the discrepancy between near-solutions of the optimality condition 0∈Sν​(x,y,z)0superscript𝑆𝜈𝑥𝑦𝑧0\in S^{\nu}(x...
1,684
6 Rates and Error Estimates Consistency furnishes guarantees about the limiting behavior of approximations, but it also can be beneficial to quantify the rate of convergence. In this section, we refine results from [53] and estimate the discrepancy between near-solutions of the optimality condition 0∈Sν​(x,y,z)0supersc...
1,690
57
Thus, S−1​(𝔹​(0,ε))superscript𝑆1𝔹0𝜀S^{-1}(\mathbb{B}(0,\varepsilon)) is the set of near-solutions of 0∈S​(x,y,z)0𝑆𝑥𝑦𝑧0\in S(x,y,z) with the tolerance now being specified by ∥⋅∥out\|\cdot\|_{\rm out}, i.e., (x¯,y¯,z¯)∈S−1​(𝔹​(0,ε))¯𝑥¯𝑦¯𝑧superscript𝑆1𝔹0𝜀(\bar{x},\bar{y},\bar{z})\in S^{-1}(\mathbb{B}(0,\var...
352
6 Rates and Error Estimates Thus, S−1​(𝔹​(0,ε))superscript𝑆1𝔹0𝜀S^{-1}(\mathbb{B}(0,\varepsilon)) is the set of near-solutions of 0∈S​(x,y,z)0𝑆𝑥𝑦𝑧0\in S(x,y,z) with the tolerance now being specified by ∥⋅∥out\|\cdot\|_{\rm out}, i.e., (x¯,y¯,z¯)∈S−1​(𝔹​(0,ε))¯𝑥¯𝑦¯𝑧superscript𝑆1𝔹0𝜀(\bar{x},\bar{y},\bar{z})...
358
58
(solution error in optimality conditions). Suppose that 0≤δν≤ρ<∞0superscript𝛿𝜈𝜌0\leq\delta^{\nu}\leq\rho<\infty and ε≥δν+exsρ⁡(gph⁡Sν;gph⁡S)𝜀superscript𝛿𝜈subscriptexs𝜌gphsuperscript𝑆𝜈gph𝑆\varepsilon\geq\delta^{\nu}+\operatorname{exs}_{\rho}(\operatorname{gph}S^{\nu};~{}\operatorname{gph}S). Then, under the no...
1,243
6 Rates and Error Estimates 6.1 Proposition (solution error in optimality conditions). Suppose that 0≤δν≤ρ<∞0superscript𝛿𝜈𝜌0\leq\delta^{\nu}\leq\rho<\infty and ε≥δν+exsρ⁡(gph⁡Sν;gph⁡S)𝜀superscript𝛿𝜈subscriptexs𝜌gphsuperscript𝑆𝜈gph𝑆\varepsilon\geq\delta^{\nu}+\operatorname{exs}_{\rho}(\operatorname{gph}S^{\nu}...
1,254
59
(estimate of excess). Suppose that ρ∈[0,∞)𝜌0\rho\in[0,\infty) and X=Xν𝑋superscript𝑋𝜈X=X^{\nu}. Let , 1 = η0νsuperscriptsubscript𝜂0𝜈\displaystyle\eta_{0}^{\nu}. , 2 = =sup{‖Fν​(x)−F​(x)‖2|x∈X∩𝔹​(0,ρ)}absentsupremumconditionalsubscriptnormsuperscript𝐹𝜈𝑥𝐹𝑥2𝑥𝑋𝔹0𝜌\displaystyle=\sup\Big{\{}\big{\|}F^{\nu}(x)-...
1,992
6 Rates and Error Estimates 6.2 Theorem (estimate of excess). Suppose that ρ∈[0,∞)𝜌0\rho\in[0,\infty) and X=Xν𝑋superscript𝑋𝜈X=X^{\nu}. Let , 1 = η0νsuperscriptsubscript𝜂0𝜈\displaystyle\eta_{0}^{\nu}. , 2 = =sup{‖Fν​(x)−F​(x)‖2|x∈X∩𝔹​(0,ρ)}absentsupremumconditionalsubscriptnormsuperscript𝐹𝜈𝑥𝐹𝑥2𝑥𝑋𝔹0𝜌\disp...
2,004
60
There are a¯i∈con​∂fiν​(x¯)subscript¯𝑎𝑖consuperscriptsubscript𝑓𝑖𝜈¯𝑥\bar{a}_{i}\in\operatorname{con}\partial f_{i}^{\nu}(\bar{x}) such that w¯−∑i=1my¯i​a¯i∈NX​(x¯)¯𝑤superscriptsubscript𝑖1𝑚subscript¯𝑦𝑖subscript¯𝑎𝑖subscript𝑁𝑋¯𝑥\bar{w}-\mathop{\sum}\nolimits_{i=1}^{m}\bar{y}_{i}\bar{a}_{i}\in N_{X}(\bar{x})...
1,881
6 Rates and Error Estimates 6.2 Theorem There are a¯i∈con​∂fiν​(x¯)subscript¯𝑎𝑖consuperscriptsubscript𝑓𝑖𝜈¯𝑥\bar{a}_{i}\in\operatorname{con}\partial f_{i}^{\nu}(\bar{x}) such that w¯−∑i=1my¯i​a¯i∈NX​(x¯)¯𝑤superscriptsubscript𝑖1𝑚subscript¯𝑦𝑖subscript¯𝑎𝑖subscript𝑁𝑋¯𝑥\bar{w}-\mathop{\sum}\nolimits_{i=1}^{m}...
1,893
61
, 1 = d​l^ρ​(C,D)=max⁡{exsρ⁡(C;D);exsρ⁡(D;C)}.𝑑subscript^𝑙𝜌𝐶𝐷subscriptexs𝜌𝐶𝐷subscriptexs𝜌𝐷𝐶d\hat{\kern-1.49994ptl}_{\rho}(C,D)=\max\big{\{}\operatorname{exs}_{\rho}(C;D);\,\operatorname{exs}_{\rho}(D;C)\big{\}}.. , 2 =
144
6 Rates and Error Estimates 6.2 Theorem , 1 = d​l^ρ​(C,D)=max⁡{exsρ⁡(C;D);exsρ⁡(D;C)}.𝑑subscript^𝑙𝜌𝐶𝐷subscriptexs𝜌𝐶𝐷subscriptexs𝜌𝐷𝐶d\hat{\kern-1.49994ptl}_{\rho}(C,D)=\max\big{\{}\operatorname{exs}_{\rho}(C;D);\,\operatorname{exs}_{\rho}(D;C)\big{\}}.. , 2 =
156
62
(approximation of subgradients). Suppose that the norm on ℝm+1superscriptℝ𝑚1\mathbb{R}^{m+1} is max⁡{‖z‖2,|α|}subscriptnorm𝑧2𝛼\max\{\|z\|_{2},|\alpha|\}, the norm on ℝm×ℝmsuperscriptℝ𝑚superscriptℝ𝑚\mathbb{R}^{m}\times\mathbb{R}^{m} is max⁡{‖z‖2,‖v‖2}subscriptnorm𝑧2subscriptnorm𝑣2\max\{\|z\|_{2},\|v\|_{2}\}, and ...
755
6 Rates and Error Estimates 6.3 Proposition (approximation of subgradients). Suppose that the norm on ℝm+1superscriptℝ𝑚1\mathbb{R}^{m+1} is max⁡{‖z‖2,|α|}subscriptnorm𝑧2𝛼\max\{\|z\|_{2},|\alpha|\}, the norm on ℝm×ℝmsuperscriptℝ𝑚superscriptℝ𝑚\mathbb{R}^{m}\times\mathbb{R}^{m} is max⁡{‖z‖2,‖v‖2}subscriptnorm𝑧2subsc...
766
63
(goal optimization; cont.). In Example 3.1, smoothing of hνsuperscriptℎ𝜈h^{\nu} causes a solution error , 1 = exsρ⁡((Sν)−1​(𝔹​(0,δν));S−1​(𝔹​(0,ε)))≤β/θνsubscriptexs𝜌superscriptsuperscript𝑆𝜈1𝔹0superscript𝛿𝜈superscript𝑆1𝔹0𝜀𝛽superscript𝜃𝜈\operatorname{exs}_{\rho}\Big{(}\,(S^{\nu})^{-1}\big{(}\mathbb{B}(0,\...
818
6 Rates and Error Estimates 6.4 Example (goal optimization; cont.). In Example 3.1, smoothing of hνsuperscriptℎ𝜈h^{\nu} causes a solution error , 1 = exsρ⁡((Sν)−1​(𝔹​(0,δν));S−1​(𝔹​(0,ε)))≤β/θνsubscriptexs𝜌superscriptsuperscript𝑆𝜈1𝔹0superscript𝛿𝜈superscript𝑆1𝔹0𝜀𝛽superscript𝜃𝜈\operatorname{exs}_{\rho}\Big...
829
64
(distributionally robust optimization; cont.). For Example 3.3, there is a constant β∈[0,∞)𝛽0\beta\in[0,\infty) such that , 1 = exsρ⁡((Sν)−1​(𝔹​(0,δν));S−1​(𝔹​(0,ε)))≤β​ανsubscriptexs𝜌superscriptsuperscript𝑆𝜈1𝔹0superscript𝛿𝜈superscript𝑆1𝔹0𝜀𝛽superscript𝛼𝜈\operatorname{exs}_{\rho}\Big{(}\,(S^{\nu})^{-1}\bi...
1,289
6 Rates and Error Estimates 6.5 Example (distributionally robust optimization; cont.). For Example 3.3, there is a constant β∈[0,∞)𝛽0\beta\in[0,\infty) such that , 1 = exsρ⁡((Sν)−1​(𝔹​(0,δν));S−1​(𝔹​(0,ε)))≤β​ανsubscriptexs𝜌superscriptsuperscript𝑆𝜈1𝔹0superscript𝛿𝜈superscript𝑆1𝔹0𝜀𝛽superscript𝛼𝜈\operatorna...
1,300
65
(augmented Lagrangian methods; cont.). For Example 3.4, we obtain that , 1 = exsρ⁡((Sν)−1​(𝔹​(0,δν));S−1​(𝔹​(0,ε)))≤βν/θν, where ​βν=(2​ρ+‖yν‖∞)​m−1,formulae-sequencesubscriptexs𝜌superscriptsuperscript𝑆𝜈1𝔹0superscript𝛿𝜈superscript𝑆1𝔹0𝜀superscript𝛽𝜈superscript𝜃𝜈 where superscript𝛽𝜈2𝜌subscriptnormsupers...
1,951
6 Rates and Error Estimates 6.6 Example (augmented Lagrangian methods; cont.). For Example 3.4, we obtain that , 1 = exsρ⁡((Sν)−1​(𝔹​(0,δν));S−1​(𝔹​(0,ε)))≤βν/θν, where ​βν=(2​ρ+‖yν‖∞)​m−1,formulae-sequencesubscriptexs𝜌superscriptsuperscript𝑆𝜈1𝔹0superscript𝛿𝜈superscript𝑆1𝔹0𝜀superscript𝛽𝜈superscript𝜃𝜈 whe...
1,962
66
(exact penalty methods). We next approach the problem in Example 3.4 using an exact penalty method. Thus, h​(z)=z1+∑i=2mι{0}​(zi)ℎ𝑧subscript𝑧1superscriptsubscript𝑖2𝑚subscript𝜄0subscript𝑧𝑖h(z)=z_{1}+\sum_{i=2}^{m}\iota_{\{0\}}(z_{i}) as before, but for θν∈[0,∞)superscript𝜃𝜈0\theta^{\nu}\in[0,\infty) we set , 1 ...
1,592
6 Rates and Error Estimates 6.7 Example (exact penalty methods). We next approach the problem in Example 3.4 using an exact penalty method. Thus, h​(z)=z1+∑i=2mι{0}​(zi)ℎ𝑧subscript𝑧1superscriptsubscript𝑖2𝑚subscript𝜄0subscript𝑧𝑖h(z)=z_{1}+\sum_{i=2}^{m}\iota_{\{0\}}(z_{i}) as before, but for θν∈[0,∞)superscript𝜃...
1,603
67
(homotopy method; cont.). In Example 3.10, we obtain for λν∈(0,1)superscript𝜆𝜈01\lambda^{\nu}\in(0,1) that , 1 = exsρ⁡((Sν)−1​(𝔹​(0,δν));S−1​(𝔹​(0,ε)))≤βν​λν, where ​βν=1+4​ρ2−(λν)2(1−λν)2,formulae-sequencesubscriptexs𝜌superscriptsuperscript𝑆𝜈1𝔹0superscript𝛿𝜈superscript𝑆1𝔹0𝜀superscript𝛽𝜈superscript𝜆𝜈 w...
1,849
6 Rates and Error Estimates 6.8 Example (homotopy method; cont.). In Example 3.10, we obtain for λν∈(0,1)superscript𝜆𝜈01\lambda^{\nu}\in(0,1) that , 1 = exsρ⁡((Sν)−1​(𝔹​(0,δν));S−1​(𝔹​(0,ε)))≤βν​λν, where ​βν=1+4​ρ2−(λν)2(1−λν)2,formulae-sequencesubscriptexs𝜌superscriptsuperscript𝑆𝜈1𝔹0superscript𝛿𝜈superscript...
1,860
0
We consider multi-level composite optimization problems where each mapping in the composition is the expectation over a family of randomly chosen smooth mappings or the sum of some finite number of smooth mappings. We present a normalized proximal approximate gradient (NPAG) method where the approximate gradients are o...
223
Multi-Level Composite Stochastic Optimization via Nested Variance Reduction Abstract We consider multi-level composite optimization problems where each mapping in the composition is the expectation over a family of randomly chosen smooth mappings or the sum of some finite number of smooth mappings. We present a normali...
237
1
composite stochastic optimization, proximal gradient method, variance reduction
12
Multi-Level Composite Stochastic Optimization via Nested Variance Reduction keywords: composite stochastic optimization, proximal gradient method, variance reduction
26
2
In this paper, we consider multi-level composite stochastic optimization problems of the form , 1 = minimizex∈𝐑d𝐄ξm​[fm,ξm​(⋯​𝐄ξ2​[f2,ξ2​(𝐄ξ1​[f1,ξ1​(x)])]​⋯)]+Ψ​(x),subscriptminimize𝑥superscript𝐑𝑑subscript𝐄subscript𝜉𝑚delimited-[]subscript𝑓𝑚subscript𝜉𝑚⋯subscript𝐄subscript𝜉2delimited-[]subscript𝑓2subscr...
1,698
Multi-Level Composite Stochastic Optimization via Nested Variance Reduction 1 Introduction In this paper, we consider multi-level composite stochastic optimization problems of the form , 1 = minimizex∈𝐑d𝐄ξm​[fm,ξm​(⋯​𝐄ξ2​[f2,ξ2​(𝐄ξ1​[f1,ξ1​(x)])]​⋯)]+Ψ​(x),subscriptminimize𝑥superscript𝐑𝑑subscript𝐄subscript𝜉𝑚d...
1,713
3
Existing work on composite stochastic optimization traces back to [12]. Several recent progresses have been made for two-level (m=2𝑚2m=2) problems, both for the general stochastic formulation (e.g., [35, 36, 39, 15]) and for the finite-sum formulation (e.g., [20, 43, 16, 41]). In order to cite specific results, we fir...
1,761
Multi-Level Composite Stochastic Optimization via Nested Variance Reduction 1 Introduction 1.1 Related work Existing work on composite stochastic optimization traces back to [12]. Several recent progresses have been made for two-level (m=2𝑚2m=2) problems, both for the general stochastic formulation (e.g., [35, 36, 39,...
1,782
4
In this paper, we propose stochastic gradient algorithms with nested variance-reduction for solving problems (1) and (2) for any m≥1𝑚1m\geq 1, and show that their sample complexity for finding x¯¯𝑥\bar{x} such that 𝐄​[‖𝒢​(x¯)‖]≤ϵ𝐄delimited-[]norm𝒢¯𝑥italic-ϵ\mathbf{E}[\|\mathcal{G}(\bar{x})\|]\leq\epsilon is O​(ϵ...
1,495
Multi-Level Composite Stochastic Optimization via Nested Variance Reduction 1 Introduction 1.2 Contributions and outline In this paper, we propose stochastic gradient algorithms with nested variance-reduction for solving problems (1) and (2) for any m≥1𝑚1m\geq 1, and show that their sample complexity for finding x¯¯𝑥...
1,517
5
In this section, we present a normalized proximal approximate gradient (NPAG) method for solving problems of form (3), which we repeat here for convenience: , 1 = minimizex∈𝐑d{Φ​(x)≜F​(x)+Ψ​(x)}.subscriptminimize𝑥superscript𝐑𝑑≜Φ𝑥𝐹𝑥Ψ𝑥\mathop{\textrm{minimize}}_{x\in\mathbf{R}^{d}}~{}\left\{\Phi(x)\triangleq F(x)...
187
Multi-Level Composite Stochastic Optimization via Nested Variance Reduction 2 Normalized proximal approximate gradient method In this section, we present a normalized proximal approximate gradient (NPAG) method for solving problems of form (3), which we repeat here for convenience: , 1 = minimizex∈𝐑d{Φ​(x)≜F​(x)+Ψ​(x)...
208
6
The functions F𝐹F, ΨΨ\Psi and ΦΦ\Phi in (5) satisfy: (a) F:𝐑d→𝐑:𝐹→superscript𝐑𝑑𝐑F:\mathbf{R}^{d}\to\mathbf{R} is differentiable and its gradient F′superscript𝐹′F^{\prime} is L𝐿L-Lipschitz continuous; (b) Ψ:𝐑d→𝐑∪{+∞}:Ψ→superscript𝐑𝑑𝐑\Psi:\mathbf{R}^{d}\to\mathbf{R}\cup\{+\infty\} is convex and lower semi...
1,515
Multi-Level Composite Stochastic Optimization via Nested Variance Reduction 2 Normalized proximal approximate gradient method Assumption 2.1. The functions F𝐹F, ΨΨ\Psi and ΦΦ\Phi in (5) satisfy: (a) F:𝐑d→𝐑:𝐹→superscript𝐑𝑑𝐑F:\mathbf{R}^{d}\to\mathbf{R} is differentiable and its gradient F′superscript𝐹′F^{\prime...
1,543
7
Here we focus on the case where the full gradient oracle F′​(⋅)superscript𝐹′⋅F^{\prime}(\cdot) is not available; instead, we can compute at each iteration t𝑡t an approximate gradient vtsuperscript𝑣𝑡v^{t}. A straightforward approach is to replace F′​(xt)superscript𝐹′superscript𝑥𝑡F^{\prime}(x^{t}) with vtsuperscri...
1,893
Multi-Level Composite Stochastic Optimization via Nested Variance Reduction 2 Normalized proximal approximate gradient method Assumption 2.1. Here we focus on the case where the full gradient oracle F′​(⋅)superscript𝐹′⋅F^{\prime}(\cdot) is not available; instead, we can compute at each iteration t𝑡t an approximate gr...
1,921
8
Next, we prove a general convergence result concerning Algorithm 1 without specifying how the approximate gradient vtsuperscript𝑣𝑡v^{t} is generated. The only condition we impose is that the Mean-Square Error (MSE), 𝐄​[‖vt−F′​(xt)‖2]𝐄delimited-[]superscriptnormsuperscript𝑣𝑡superscript𝐹′superscript𝑥𝑡2\mathbf{E}...
154
Multi-Level Composite Stochastic Optimization via Nested Variance Reduction 2 Normalized proximal approximate gradient method Assumption 2.1. Next, we prove a general convergence result concerning Algorithm 1 without specifying how the approximate gradient vtsuperscript𝑣𝑡v^{t} is generated. The only condition we impo...
182
9
Suppose Assumption 2.1 hold. Then the sequence {xt}superscript𝑥𝑡\{x^{t}\} generated by Algorithm 1 satisfies , 1 = Φ​(xt+1)≤Φ​(xt)−(γt/η−L​γt2)​∥x~t+1−xt∥2+12​L​∥F′​(xt)−vt∥2,∀t≥0.formulae-sequenceΦsuperscript𝑥𝑡1Φsuperscript𝑥𝑡subscript𝛾𝑡𝜂𝐿superscriptsubscript𝛾𝑡2superscriptdelimited-∥∥superscript~𝑥𝑡1supers...
351
Multi-Level Composite Stochastic Optimization via Nested Variance Reduction 2 Normalized proximal approximate gradient method Lemma 2.2. Suppose Assumption 2.1 hold. Then the sequence {xt}superscript𝑥𝑡\{x^{t}\} generated by Algorithm 1 satisfies , 1 = Φ​(xt+1)≤Φ​(xt)−(γt/η−L​γt2)​∥x~t+1−xt∥2+12​L​∥F′​(xt)−vt∥2,∀t≥0.f...
378
10
According to the update rule for xt+1superscript𝑥𝑡1x^{t+1}, we have , 1 = Φ​(xt+1)Φsuperscript𝑥𝑡1\displaystyle\Phi(x^{t+1}). , 2 = =\displaystyle=. , 3 = F​(xt+γt​(x~t+1−xt))+Ψ​(xt+γt​(x~t+1−xt))𝐹superscript𝑥𝑡subscript𝛾𝑡superscript~𝑥𝑡1superscript𝑥𝑡Ψsuperscript𝑥𝑡subscript𝛾𝑡superscript~𝑥𝑡1superscript𝑥...
1,775
Multi-Level Composite Stochastic Optimization via Nested Variance Reduction 2 Normalized proximal approximate gradient method Proof. According to the update rule for xt+1superscript𝑥𝑡1x^{t+1}, we have , 1 = Φ​(xt+1)Φsuperscript𝑥𝑡1\displaystyle\Phi(x^{t+1}). , 2 = =\displaystyle=. , 3 = F​(xt+γt​(x~t+1−xt))+Ψ​(xt+γt...
1,798
11
, 1 = ⟨vt,x~t+1−xt⟩+Ψ​(x~t+1)−Ψ​(xt)≤−1η​∥x~t+1−xt∥2.superscript𝑣𝑡superscript~𝑥𝑡1superscript𝑥𝑡Ψsuperscript~𝑥𝑡1Ψsuperscript𝑥𝑡1𝜂superscriptdelimited-∥∥superscript~𝑥𝑡1superscript𝑥𝑡2\left\langle v^{t},\tilde{x}^{t+1}-x^{t}\right\rangle+\Psi(\tilde{x}^{t+1})-\Psi(x^{t})~{}\leq~{}-\frac{1}{\eta}\bigl{\|}\tilde...
1,247
Multi-Level Composite Stochastic Optimization via Nested Variance Reduction 2 Normalized proximal approximate gradient method Proof. , 1 = ⟨vt,x~t+1−xt⟩+Ψ​(x~t+1)−Ψ​(xt)≤−1η​∥x~t+1−xt∥2.superscript𝑣𝑡superscript~𝑥𝑡1superscript𝑥𝑡Ψsuperscript~𝑥𝑡1Ψsuperscript𝑥𝑡1𝜂superscriptdelimited-∥∥superscript~𝑥𝑡1superscrip...
1,270
12
Suppose Assumption 2.1 hold and we set η=1/2​L𝜂12𝐿\eta=1/2L in Algorithm 1. If it holds that , 1 = 𝐄​[‖vt−F′​(xt)‖2]≤ϵt2,∀t≥0,formulae-sequence𝐄delimited-[]superscriptnormsuperscript𝑣𝑡superscript𝐹′superscript𝑥𝑡2superscriptsubscriptitalic-ϵ𝑡2for-all𝑡0\mathbf{E}\left[\|v^{t}-F^{\prime}(x^{t})\|^{2}\right]\leq\...
538
Multi-Level Composite Stochastic Optimization via Nested Variance Reduction 2 Normalized proximal approximate gradient method Theorem 2.3. Suppose Assumption 2.1 hold and we set η=1/2​L𝜂12𝐿\eta=1/2L in Algorithm 1. If it holds that , 1 = 𝐄​[‖vt−F′​(xt)‖2]≤ϵt2,∀t≥0,formulae-sequence𝐄delimited-[]superscriptnormsupers...
566
13
Using Lemma 2.2, the assumption (12), and the choice η=1/2​L𝜂12𝐿\eta=1/2L, we obtain , 1 = 𝐄​[Φ​(xt+1)]𝐄delimited-[]Φsuperscript𝑥𝑡1\displaystyle\mathbf{E}\left[\Phi(x^{t+1})\right]. , 2 = ≤\displaystyle\leq. , 3 = 𝐄​[Φ​(xt)]−𝐄​[(γt/η−L​γt2)​∥x~t+1−xt∥2]+ϵt22​L𝐄delimited-[]Φsuperscript𝑥𝑡𝐄delimited-[]subscrip...
1,951
Multi-Level Composite Stochastic Optimization via Nested Variance Reduction 2 Normalized proximal approximate gradient method Proof. Using Lemma 2.2, the assumption (12), and the choice η=1/2​L𝜂12𝐿\eta=1/2L, we obtain , 1 = 𝐄​[Φ​(xt+1)]𝐄delimited-[]Φsuperscript𝑥𝑡1\displaystyle\mathbf{E}\left[\Phi(x^{t+1})\right]....
1,974
14
, 1 = 𝐄​[‖vt−F′​(xt)‖]≤𝐄​[‖vt−F′​(xt)‖2]≤ϵt.𝐄delimited-[]normsuperscript𝑣𝑡superscript𝐹′superscript𝑥𝑡𝐄delimited-[]superscriptnormsuperscript𝑣𝑡superscript𝐹′superscript𝑥𝑡2subscriptitalic-ϵ𝑡\mathbf{E}\left[\|v^{t}-F^{\prime}(x^{t})\|\right]\leq\sqrt{\mathbf{E}\left[\|v^{t}-F^{\prime}(x^{t})\|^{2}\right]}\leq...
1,516
Multi-Level Composite Stochastic Optimization via Nested Variance Reduction 2 Normalized proximal approximate gradient method Proof. , 1 = 𝐄​[‖vt−F′​(xt)‖]≤𝐄​[‖vt−F′​(xt)‖2]≤ϵt.𝐄delimited-[]normsuperscript𝑣𝑡superscript𝐹′superscript𝑥𝑡𝐄delimited-[]superscriptnormsuperscript𝑣𝑡superscript𝐹′superscript𝑥𝑡2subsc...
1,539
15
, 1 = 𝐄​[∥𝒢​(x¯)∥]=∑k=0T−1ϵk∑t=0T−1ϵt⋅𝐄​[∥𝒢​(xt)∥]≤4​L​(Φ​(x0)−Φ∗)∑t=0T−1ϵt+4​∑t=0T−1ϵt2∑t=0T−1ϵt.𝐄delimited-[]delimited-∥∥𝒢¯𝑥superscriptsubscript𝑘0𝑇1⋅subscriptitalic-ϵ𝑘superscriptsubscript𝑡0𝑇1subscriptitalic-ϵ𝑡𝐄delimited-[]delimited-∥∥𝒢superscript𝑥𝑡4𝐿Φsuperscript𝑥0subscriptΦsuperscriptsubscript𝑡0𝑇...
481
Multi-Level Composite Stochastic Optimization via Nested Variance Reduction 2 Normalized proximal approximate gradient method Proof. , 1 = 𝐄​[∥𝒢​(x¯)∥]=∑k=0T−1ϵk∑t=0T−1ϵt⋅𝐄​[∥𝒢​(xt)∥]≤4​L​(Φ​(x0)−Φ∗)∑t=0T−1ϵt+4​∑t=0T−1ϵt2∑t=0T−1ϵt.𝐄delimited-[]delimited-∥∥𝒢¯𝑥superscriptsubscript𝑘0𝑇1⋅subscriptitalic-ϵ𝑘superscr...
504
16
In the simplest case, one can set ϵt≡ϵsubscriptitalic-ϵ𝑡italic-ϵ\epsilon_{t}\equiv\epsilon for all t=0,1,…,T−1𝑡01…𝑇1t=0,1,...,T-1 and , 1 = T≥4​L​(Φ​(x0)−Φ∗)​ϵ−2.𝑇4𝐿Φsuperscript𝑥0subscriptΦsuperscriptitalic-ϵ2T\geq 4L\left(\Phi(x^{0})-\Phi_{*}\right)\epsilon^{-2}.. , 2 = Then Theorem 2.3 directly implies , 1 = �...
874
Multi-Level Composite Stochastic Optimization via Nested Variance Reduction 2 Normalized proximal approximate gradient method Remark 2.4. In the simplest case, one can set ϵt≡ϵsubscriptitalic-ϵ𝑡italic-ϵ\epsilon_{t}\equiv\epsilon for all t=0,1,…,T−1𝑡01…𝑇1t=0,1,...,T-1 and , 1 = T≥4​L​(Φ​(x0)−Φ∗)​ϵ−2.𝑇4𝐿Φsuperscript...
901
17
In this section, we discuss stochastic variance-reduction techniques for smooth nonconvex optimization. In order to prepare for the multi-level compositional case, we proceed with a general framework of constructing stochastic estimators that satisfies (12) for Lipschitz-continuous vector or matrix mappings. To simplif...
1,976
Multi-Level Composite Stochastic Optimization via Nested Variance Reduction 3 A general framework of stochastic variance reduction In this section, we discuss stochastic variance-reduction techniques for smooth nonconvex optimization. In order to prepare for the multi-level compositional case, we proceed with a general...
1,997
18
However, if the desired accuracy ϵitalic-ϵ\epsilon is smaller than the uncertainty level σ𝜎\sigma, then condition (12) is not satisfied and we cannot use Theorem 2.3 directly. A common remedy is to use mini-batches; i.e., at each iteration of the algorithm, we randomly pick a subset ℬtsubscriptℬ𝑡\mathcal{B}_{t} of ξ�...
1,527
Multi-Level Composite Stochastic Optimization via Nested Variance Reduction 3 A general framework of stochastic variance reduction However, if the desired accuracy ϵitalic-ϵ\epsilon is smaller than the uncertainty level σ𝜎\sigma, then condition (12) is not satisfied and we cannot use Theorem 2.3 directly. A common rem...
1,548
19
In order to construct τ𝜏\tau consecutive estimates {v0,…,vτ−1}superscript𝑣0…superscript𝑣𝜏1\{v^{0},\ldots,v^{\tau-1}\}, this estimator uses the mini-batch estimator (20) for v0superscript𝑣0v^{0}, and then constructs v1superscript𝑣1v^{1} through vτ−1superscript𝑣𝜏1v^{\tau-1} using a recursion: , 1 = v0=ϕℬ0​(x0),vt...
452
Multi-Level Composite Stochastic Optimization via Nested Variance Reduction 3 A general framework of stochastic variance reduction 3.1 SARAH/SPIDER estimator for stochastic optimization In order to construct τ𝜏\tau consecutive estimates {v0,…,vτ−1}superscript𝑣0…superscript𝑣𝜏1\{v^{0},\ldots,v^{\tau-1}\}, this estima...
486
20
[13, Lemma 1] Suppose the random mappings ϕξsubscriptitalic-ϕ𝜉\phi_{\xi} satisfy (18). Then the MSE of the estimator in (22) can be bounded as , 1 = 𝐄​[∥vt−ϕ¯​(xt)∥2]≤𝐄​[∥v0−ϕ¯​(x0)∥2]+∑r=1tL2|ℬr|​𝐄​[‖xr−xr−1‖2].𝐄delimited-[]superscriptdelimited-∥∥superscript𝑣𝑡¯italic-ϕsuperscript𝑥𝑡2𝐄delimited-[]superscriptde...
1,896
Multi-Level Composite Stochastic Optimization via Nested Variance Reduction 3 A general framework of stochastic variance reduction 3.1 SARAH/SPIDER estimator for stochastic optimization Lemma 3.1. [13, Lemma 1] Suppose the random mappings ϕξsubscriptitalic-ϕ𝜉\phi_{\xi} satisfy (18). Then the MSE of the estimator in (2...
1,936
21
, 1 = b⋆=τ⋆/2=σ/ϵ,ρ​(τ⋆)=2​ϵ/σ.formulae-sequencesubscript𝑏⋆subscript𝜏⋆2𝜎italic-ϵ𝜌subscript𝜏⋆2italic-ϵ𝜎b_{\star}=\tau_{\star}/2=\sigma/\epsilon,\qquad\rho(\tau_{\star})=2\,\epsilon/\sigma.. , 2 = . , 3 = (26) Therefore, significant reduction in sample complexity can be expected when ϵ≪σmuch-less-thanitalic-ϵ𝜎\eps...
1,043
Multi-Level Composite Stochastic Optimization via Nested Variance Reduction 3 A general framework of stochastic variance reduction 3.1 SARAH/SPIDER estimator for stochastic optimization Lemma 3.1. , 1 = b⋆=τ⋆/2=σ/ϵ,ρ​(τ⋆)=2​ϵ/σ.formulae-sequencesubscript𝑏⋆subscript𝜏⋆2𝜎italic-ϵ𝜌subscript𝜏⋆2italic-ϵ𝜎b_{\star}=\tau_...
1,083
22
Consider problem (5) with F​(x)=𝐄ξ​[fξ​(x)]𝐹𝑥subscript𝐄𝜉delimited-[]subscript𝑓𝜉𝑥F(x)=\mathbf{E}_{\xi}[f_{\xi}(x)]. Suppose Assumption 2.1 holds and the gradient mapping ϕξ≡fξ′subscriptitalic-ϕ𝜉subscriptsuperscript𝑓′𝜉\phi_{\xi}\equiv f^{\prime}_{\xi} satisfies (18) and (19) on dom​ΨdomΨ\mathrm{dom\,}\Psi (ins...
613
Multi-Level Composite Stochastic Optimization via Nested Variance Reduction 3 A general framework of stochastic variance reduction 3.1 SARAH/SPIDER estimator for stochastic optimization Corollary 3.2. Consider problem (5) with F​(x)=𝐄ξ​[fξ​(x)]𝐹𝑥subscript𝐄𝜉delimited-[]subscript𝑓𝜉𝑥F(x)=\mathbf{E}_{\xi}[f_{\xi}(x...
655
23
In Algorithm 2, the step lengths satisfy the same bound in (9), i.e., ‖xt+1−xt‖≤η​ϵ=ϵ/2​Lnormsuperscript𝑥𝑡1superscript𝑥𝑡𝜂italic-ϵitalic-ϵ2𝐿\|x^{t+1}-x^{t}\|\leq\eta\epsilon=\epsilon/2L. From the analysis following (25), the parameters in (28) guarantee (12). Therefore we can apply Theorem 2.3 and Remark 2.4 to co...
1,010
Multi-Level Composite Stochastic Optimization via Nested Variance Reduction 3 A general framework of stochastic variance reduction 3.1 SARAH/SPIDER estimator for stochastic optimization Proof. In Algorithm 2, the step lengths satisfy the same bound in (9), i.e., ‖xt+1−xt‖≤η​ϵ=ϵ/2​Lnormsuperscript𝑥𝑡1superscript𝑥𝑡𝜂i...
1,046
24
Now consider a finite number of mappings ϕi:𝐑d→𝐑p×q:subscriptitalic-ϕ𝑖→superscript𝐑𝑑superscript𝐑𝑝𝑞\phi_{i}:\mathbf{R}^{d}\to\mathbf{R}^{p\times q} for i=1,…,n𝑖1…𝑛i=1,\ldots,n, and define ϕ¯​(x)=(1/n)​∑i=1nϕi​(x)¯italic-ϕ𝑥1𝑛superscriptsubscript𝑖1𝑛subscriptitalic-ϕ𝑖𝑥\bar{\phi}(x)=(1/n)\sum_{i=1}^{n}\phi_{...
1,789
Multi-Level Composite Stochastic Optimization via Nested Variance Reduction 3 A general framework of stochastic variance reduction 3.2 SARAH/SPIDER estimator for finite-sum optimization Now consider a finite number of mappings ϕi:𝐑d→𝐑p×q:subscriptitalic-ϕ𝑖→superscript𝐑𝑑superscript𝐑𝑝𝑞\phi_{i}:\mathbf{R}^{d}\to\m...
1,824
25
Consider problem (5) with F​(x)=(1/n)​∑i=1nfi​(x)𝐹𝑥1𝑛superscriptsubscript𝑖1𝑛subscript𝑓𝑖𝑥F(x)=(1/n)\sum_{i=1}^{n}f_{i}(x). Suppose Assumption 2.1 holds and the gradient fi′subscriptsuperscript𝑓′𝑖f^{\prime}_{i} satisfies (18) on dom​ΨdomΨ\mathrm{dom\,}\Psi. If we set the parameters in Algorithm 2 as , 1 = η=1/2...
516
Multi-Level Composite Stochastic Optimization via Nested Variance Reduction 3 A general framework of stochastic variance reduction 3.2 SARAH/SPIDER estimator for finite-sum optimization Corollary 3.3. Consider problem (5) with F​(x)=(1/n)​∑i=1nfi​(x)𝐹𝑥1𝑛superscriptsubscript𝑖1𝑛subscript𝑓𝑖𝑥F(x)=(1/n)\sum_{i=1}^{n...
559
26
As a general framework, NPAG may also incorporate the SVRG estimator proposed in [18]. We illustrate how SVRG estimator can be applied to construct τ𝜏\tau consecutive estimates {v0,…,vτ−1}superscript𝑣0…superscript𝑣𝜏1\{v^{0},...,v^{\tau-1}\} that satisfy (21) for stochastic optimization (the expectation case). For t...
1,453
Multi-Level Composite Stochastic Optimization via Nested Variance Reduction 3 A general framework of stochastic variance reduction 3.3 SVRG estimator for stochastic optimization As a general framework, NPAG may also incorporate the SVRG estimator proposed in [18]. We illustrate how SVRG estimator can be applied to cons...
1,484
27
We can minimize the ratio between the above quantity and τ​σ2/ϵ2𝜏superscript𝜎2superscriptitalic-ϵ2\tau\sigma^{2}/\epsilon^{2} (of the naive mini-batch scheme) by choosing τ=(2​σ/ϵ)2/3𝜏superscript2𝜎italic-ϵ23\tau=(\sqrt{2}\sigma/\epsilon)^{2/3} and total sample complexity per epoch becomes O​(σ2/ϵ2)𝑂superscript𝜎2s...
558
Multi-Level Composite Stochastic Optimization via Nested Variance Reduction 3 A general framework of stochastic variance reduction 3.3 SVRG estimator for stochastic optimization We can minimize the ratio between the above quantity and τ​σ2/ϵ2𝜏superscript𝜎2superscriptitalic-ϵ2\tau\sigma^{2}/\epsilon^{2} (of the naive ...
589
28
In this section, we discussion how the SAGA estimator proposed in [8] can be incorporated into our NPAG framework. Note that the SAGA estimator is only meaningful for the finite-sum case (sampling with replacement). The SAGA estimator can be written as , 1 = v0superscript𝑣0\displaystyle v^{0}. , 2 = =\displaystyle=. ,...
1,536
Multi-Level Composite Stochastic Optimization via Nested Variance Reduction 3 A general framework of stochastic variance reduction 3.4 SAGA estimator for finite-sum optimization In this section, we discussion how the SAGA estimator proposed in [8] can be incorporated into our NPAG framework. Note that the SAGA estimato...
1,569
29
When we restrict the step lengths by ‖xr−xr−1‖≤δnormsuperscript𝑥𝑟superscript𝑥𝑟1𝛿\|x^{r}-x^{r-1}\|\leq\delta, then we have ‖xt−xr‖≤|t−r|​δnormsuperscript𝑥𝑡superscript𝑥𝑟𝑡𝑟𝛿\|x^{t}-x^{r}\|\leq|t-r|\delta for all t,r𝑡𝑟t,r. Note that all the batches ℬrsubscriptℬ𝑟\mathcal{B}_{r} are independently and uniformly...
1,695
Multi-Level Composite Stochastic Optimization via Nested Variance Reduction 3 A general framework of stochastic variance reduction 3.4 SAGA estimator for finite-sum optimization When we restrict the step lengths by ‖xr−xr−1‖≤δnormsuperscript𝑥𝑟superscript𝑥𝑟1𝛿\|x^{r}-x^{r-1}\|\leq\delta, then we have ‖xt−xr‖≤|t−r|​δ...
1,728
30
In this section, we present results for the general composite optimization problems (1) and (2) with m≥2𝑚2m\geq 2. To provide convergence and complexity result, we only need to construct a gradient estimator vtsuperscript𝑣𝑡v^{t} that is sufficiently close to F′​(xt)superscript𝐹′superscript𝑥𝑡F^{\prime}(x^{t}). For...
1,768
Multi-Level Composite Stochastic Optimization via Nested Variance Reduction 4 Multi-level nested SPIDER In this section, we present results for the general composite optimization problems (1) and (2) with m≥2𝑚2m\geq 2. To provide convergence and complexity result, we only need to construct a gradient estimator vtsuper...
1,787
31
For the functions and mappings appearing in (1) and (2), we assume: (a) For each i=1,…,m𝑖1…𝑚i=1,\ldots,m and each realization of ξisubscript𝜉𝑖\xi_{i}, mapping fi,ξi:𝐑di−1→𝐑di:subscript𝑓𝑖subscript𝜉𝑖→superscript𝐑subscript𝑑𝑖1superscript𝐑subscript𝑑𝑖f_{i,\xi_{i}}\!:\mathbf{R}^{d_{i-1}}\!\to\!\mathbf{R}^{d_{...
907
Multi-Level Composite Stochastic Optimization via Nested Variance Reduction 4 Multi-level nested SPIDER Assumption 4.1. For the functions and mappings appearing in (1) and (2), we assume: (a) For each i=1,…,m𝑖1…𝑚i=1,\ldots,m and each realization of ξisubscript𝜉𝑖\xi_{i}, mapping fi,ξi:𝐑di−1→𝐑di:subscript𝑓𝑖subsc...
933
32
Suppose Assumption 4.1.(a) holds. Then the composite function F𝐹F and its gradient F′superscript𝐹′F^{\prime} are Lipschitz continuous, with respective Lipschitz constants , 1 = ℓF=∏i=1mℓi,LF=∑i=1mLi​(∏r=1i−1ℓr2)​(∏r=i+1mℓr),formulae-sequencesubscriptℓ𝐹superscriptsubscriptproduct𝑖1𝑚subscriptℓ𝑖subscript𝐿𝐹superscr...
1,175
Multi-Level Composite Stochastic Optimization via Nested Variance Reduction 4 Multi-level nested SPIDER Lemma 4.2. Suppose Assumption 4.1.(a) holds. Then the composite function F𝐹F and its gradient F′superscript𝐹′F^{\prime} are Lipschitz continuous, with respective Lipschitz constants , 1 = ℓF=∏i=1mℓi,LF=∑i=1mLi​(∏r=...
1,200
33
Suppose Assumption 4.1 holds. In Algorithm 3, if we set η=12​LF𝜂12subscript𝐿𝐹\eta=\frac{1}{2L_{F}}, τk=ℓF2​m​ϵksubscript𝜏𝑘subscriptℓ𝐹2𝑚subscriptitalic-ϵ𝑘\tau_{k}=\frac{\ell_{F}}{2m\epsilon_{k}} and , 1 = Bik=12​m​(m+1)​σF2ϵk2,bik=6​(m+1)​ℓFϵk,i=1,…,m,formulae-sequencesuperscriptsubscript𝐵𝑖𝑘12𝑚𝑚1superscript...
752
Multi-Level Composite Stochastic Optimization via Nested Variance Reduction 4 Multi-level nested SPIDER Lemma 4.3. Suppose Assumption 4.1 holds. In Algorithm 3, if we set η=12​LF𝜂12subscript𝐿𝐹\eta=\frac{1}{2L_{F}}, τk=ℓF2​m​ϵksubscript𝜏𝑘subscriptℓ𝐹2𝑚subscriptitalic-ϵ𝑘\tau_{k}=\frac{\ell_{F}}{2m\epsilon_{k}} and...
777
34
First, we bound ‖vt−F′​(xt)‖2superscriptnormsuperscript𝑣𝑡superscript𝐹′superscript𝑥𝑡2\|v^{t}-F^{\prime}(x^{t})\|^{2} in terms of the approximation errors of the individual estimators in (33) and (34). Denoting Fi​(xt)subscript𝐹𝑖superscript𝑥𝑡F_{i}(x^{t}) by Fitsuperscriptsubscript𝐹𝑖𝑡F_{i}^{t} for simplicity, ...
930
Multi-Level Composite Stochastic Optimization via Nested Variance Reduction 4 Multi-level nested SPIDER Proof. First, we bound ‖vt−F′​(xt)‖2superscriptnormsuperscript𝑣𝑡superscript𝐹′superscript𝑥𝑡2\|v^{t}-F^{\prime}(x^{t})\|^{2} in terms of the approximation errors of the individual estimators in (33) and (34). Deno...
951
35
, 1 = . , 2 = . , 3 = 𝐄​[‖vt−F′​(xt)‖2]𝐄delimited-[]superscriptnormsuperscript𝑣𝑡superscript𝐹′superscript𝑥𝑡2\displaystyle\qquad\mathbf{E}\bigl{[}\|v^{t}-F^{\prime}(x^{t})\|^{2}\bigr{]}. , 4 = . , 1 = . , 2 = ≤\displaystyle\leq. , 3 = (2m−1)(𝐄[∥[z1t]T[z2t]T⋯[zm−1t]Tzmt−[f1′(xt)]T[z2t]T⋯[zm−1t]Tzmt∥2]\displaystyle...
1,871
Multi-Level Composite Stochastic Optimization via Nested Variance Reduction 4 Multi-level nested SPIDER Proof. , 1 = . , 2 = . , 3 = 𝐄​[‖vt−F′​(xt)‖2]𝐄delimited-[]superscriptnormsuperscript𝑣𝑡superscript𝐹′superscript𝑥𝑡2\displaystyle\qquad\mathbf{E}\bigl{[}\|v^{t}-F^{\prime}(x^{t})\|^{2}\bigr{]}. , 4 = . , 1 = . ,...
1,892
36
+𝐄[∥[f1′(xt)]T⋯[fm−1′(Fm−2t)]Tfm′(ym−1t)−[f1′(xt)]T⋯[fm−1′(Fm−2t)]Tfm′(Fm−1t)∥2])\displaystyle+\mathbf{E}\!\left[\big{\|}[f_{1}^{\prime}(x^{t})]^{T}\!\!\cdots\![f^{\prime}_{m\!-\!1}(F_{m\!-\!2}^{t})]^{T}\!f^{\prime}_{m}(y_{m\!-\!1}^{t})\!-\![f_{1}^{\prime}(x^{t})]^{T}\!\!\cdots\![f^{\prime}_{m\!-\!1}(F_{m\!-\!2}^{t})]...
1,868
Multi-Level Composite Stochastic Optimization via Nested Variance Reduction 4 Multi-level nested SPIDER Proof. +𝐄[∥[f1′(xt)]T⋯[fm−1′(Fm−2t)]Tfm′(ym−1t)−[f1′(xt)]T⋯[fm−1′(Fm−2t)]Tfm′(Fm−1t)∥2])\displaystyle+\mathbf{E}\!\left[\big{\|}[f_{1}^{\prime}(x^{t})]^{T}\!\!\cdots\![f^{\prime}_{m\!-\!1}(F_{m\!-\!2}^{t})]^{T}\!f^{...
1,889
37
Step 1: Bounding the temporal differences ‖yit−yit−1‖normsuperscriptsubscript𝑦𝑖𝑡superscriptsubscript𝑦𝑖𝑡1\|y_{i}^{t}-y_{i}^{t-1}\|. For i=0𝑖0i=0, we have y0t=xtsuperscriptsubscript𝑦0𝑡superscript𝑥𝑡y_{0}^{t}=x^{t} and from (9), ‖xt−xt−1‖≤η​ϵ=ϵ/2​LF≤ϵ/LFnormsuperscript𝑥𝑡superscript𝑥𝑡1𝜂italic-ϵitalic-ϵ2subsc...
1,921
Multi-Level Composite Stochastic Optimization via Nested Variance Reduction 4 Multi-level nested SPIDER Proof. Step 1: Bounding the temporal differences ‖yit−yit−1‖normsuperscriptsubscript𝑦𝑖𝑡superscriptsubscript𝑦𝑖𝑡1\|y_{i}^{t}-y_{i}^{t-1}\|. For i=0𝑖0i=0, we have y0t=xtsuperscriptsubscript𝑦0𝑡superscript𝑥𝑡y_{...
1,942
38
, 1 = ‖zit‖≤‖zi0‖+t⋅(∏r=1i−1ℓr)​Li​ϵLF≤ℓi+τ⋅Li​(∏r=1i−1ℓr)​ϵLF.normsuperscriptsubscript𝑧𝑖𝑡normsuperscriptsubscript𝑧𝑖0⋅𝑡superscriptsubscriptproduct𝑟1𝑖1subscriptℓ𝑟subscript𝐿𝑖italic-ϵsubscript𝐿𝐹subscriptℓ𝑖⋅𝜏subscript𝐿𝑖superscriptsubscriptproduct𝑟1𝑖1subscriptℓ𝑟italic-ϵsubscript𝐿𝐹\|z_{i}^{t}\|\leq\|z_{...
1,684
Multi-Level Composite Stochastic Optimization via Nested Variance Reduction 4 Multi-level nested SPIDER Proof. , 1 = ‖zit‖≤‖zi0‖+t⋅(∏r=1i−1ℓr)​Li​ϵLF≤ℓi+τ⋅Li​(∏r=1i−1ℓr)​ϵLF.normsuperscriptsubscript𝑧𝑖𝑡normsuperscriptsubscript𝑧𝑖0⋅𝑡superscriptsubscriptproduct𝑟1𝑖1subscriptℓ𝑟subscript𝐿𝑖italic-ϵsubscript𝐿𝐹subsc...
1,705
39
, 1 = 𝐄​[‖yit−Fi​(xt)‖2]≤i⋅[∑r=1i(∏j=r+1iℓj2)​δr2Sr]+i⋅(∏j=1iℓj2)​[∑r=1iτ​ϵ2sr​LF2].𝐄delimited-[]superscriptnormsuperscriptsubscript𝑦𝑖𝑡subscript𝐹𝑖superscript𝑥𝑡2⋅𝑖delimited-[]superscriptsubscript𝑟1𝑖superscriptsubscriptproduct𝑗𝑟1𝑖superscriptsubscriptℓ𝑗2superscriptsubscript𝛿𝑟2subscript𝑆𝑟⋅𝑖superscripts...
665
Multi-Level Composite Stochastic Optimization via Nested Variance Reduction 4 Multi-level nested SPIDER Proof. , 1 = 𝐄​[‖yit−Fi​(xt)‖2]≤i⋅[∑r=1i(∏j=r+1iℓj2)​δr2Sr]+i⋅(∏j=1iℓj2)​[∑r=1iτ​ϵ2sr​LF2].𝐄delimited-[]superscriptnormsuperscriptsubscript𝑦𝑖𝑡subscript𝐹𝑖superscript𝑥𝑡2⋅𝑖delimited-[]superscriptsubscript𝑟1𝑖...
686
40
, 1 = . , 2 = . , 3 = 𝐄​[‖yk+1t−Fk+1​(xt)‖2]𝐄delimited-[]superscriptnormsuperscriptsubscript𝑦𝑘1𝑡subscript𝐹𝑘1superscript𝑥𝑡2\displaystyle\mathbf{E}\bigl{[}\|y_{k+1}^{t}-F_{k+1}(x^{t})\|^{2}\bigr{]}. , 4 = . , 1 = . , 2 = ≤\displaystyle\leq. , 3 = (1+k)​𝐄​[‖yk+1t−fk+1​(ykt)‖2]+(1+k−1)​𝐄​[‖fk+1​(ykt)−fk+1​(Fk​(x...
1,926
Multi-Level Composite Stochastic Optimization via Nested Variance Reduction 4 Multi-level nested SPIDER Proof. , 1 = . , 2 = . , 3 = 𝐄​[‖yk+1t−Fk+1​(xt)‖2]𝐄delimited-[]superscriptnormsuperscriptsubscript𝑦𝑘1𝑡subscript𝐹𝑘1superscript𝑥𝑡2\displaystyle\mathbf{E}\bigl{[}\|y_{k+1}^{t}-F_{k+1}(x^{t})\|^{2}\bigr{]}. , 4...
1,947
41
, 1 = τ=ℓF2​m​ϵ=∏j=1mℓj2​m​ϵ,𝜏subscriptℓ𝐹2𝑚italic-ϵsuperscriptsubscriptproduct𝑗1𝑚subscriptℓ𝑗2𝑚italic-ϵ\tau=\frac{\ell_{F}}{2m\epsilon}=\frac{\mathop{\textstyle\prod}_{j=1}^{m}\ell_{j}}{2m\epsilon},. , 2 = we have for r=1,…,m𝑟1…𝑚r=1,\ldots,m and all t≥0𝑡0t\geq 0, , 1 = ‖zrt‖≤ℓr+τ​Lr​∏j=1r−1ℓj​ϵLF=ℓr​(1+Lr​(∏j...
1,906
Multi-Level Composite Stochastic Optimization via Nested Variance Reduction 4 Multi-level nested SPIDER Proof. , 1 = τ=ℓF2​m​ϵ=∏j=1mℓj2​m​ϵ,𝜏subscriptℓ𝐹2𝑚italic-ϵsuperscriptsubscriptproduct𝑗1𝑚subscriptℓ𝑗2𝑚italic-ϵ\tau=\frac{\ell_{F}}{2m\epsilon}=\frac{\mathop{\textstyle\prod}_{j=1}^{m}\ell_{j}}{2m\epsilon},. , 2...
1,927
42
For 𝒯1subscript𝒯1\mathcal{T}_{1}, we use (43) and set Bi=Bsubscript𝐵𝑖𝐵B_{i}=B and bi=bsubscript𝑏𝑖𝑏b_{i}=b for all i=1,…,m𝑖1…𝑚i=1,\ldots,m, which yields , 1 = 𝒯1subscript𝒯1\displaystyle\mathcal{T}_{1}. , 2 = ≤\displaystyle\leq. , 3 = 6​m​∑i=1m∏r≠iℓr2⋅(σi2Bi+Li2​(∏r=1i−1ℓr2)​τ​ϵ2bi​LF2)6𝑚superscriptsubscript...
1,565
Multi-Level Composite Stochastic Optimization via Nested Variance Reduction 4 Multi-level nested SPIDER Proof. For 𝒯1subscript𝒯1\mathcal{T}_{1}, we use (43) and set Bi=Bsubscript𝐵𝑖𝐵B_{i}=B and bi=bsubscript𝑏𝑖𝑏b_{i}=b for all i=1,…,m𝑖1…𝑚i=1,\ldots,m, which yields , 1 = 𝒯1subscript𝒯1\displaystyle\mathcal{T}_{...
1,586
43
, 1 = 𝒯2subscript𝒯2\displaystyle\mathcal{T}_{2}. , 2 = ≤6​m​∑i=1m−1(∏r≠i+1ℓr2)​Li+12⋅i⋅[∑r=1iδr2Sr​(∏j=r+1iℓj2)+(∏j=1iℓj2)​∑r=1iτ​ϵ2sr​LF2]absent6𝑚superscriptsubscript𝑖1𝑚1⋅subscriptproduct𝑟𝑖1superscriptsubscriptℓ𝑟2superscriptsubscript𝐿𝑖12𝑖delimited-[]superscriptsubscript𝑟1𝑖superscriptsubscript𝛿𝑟2subscrip...
1,816
Multi-Level Composite Stochastic Optimization via Nested Variance Reduction 4 Multi-level nested SPIDER Proof. , 1 = 𝒯2subscript𝒯2\displaystyle\mathcal{T}_{2}. , 2 = ≤6​m​∑i=1m−1(∏r≠i+1ℓr2)​Li+12⋅i⋅[∑r=1iδr2Sr​(∏j=r+1iℓj2)+(∏j=1iℓj2)​∑r=1iτ​ϵ2sr​LF2]absent6𝑚superscriptsubscript𝑖1𝑚1⋅subscriptproduct𝑟𝑖1superscript...
1,837
44
=6​m2​∑r=1m−1δr2Sr​∏j=1rℓr2​∑i=rm−1(∏j=i+1mℓj2⋅Li+12⋅∏j=1iℓj4)+6​m2​∑r=1m−1τ​ϵ2sr​LF2​∑i=rm−1(∏j=1iℓj4⋅Li+12⋅∏j=i+1mℓj2)absent6superscript𝑚2superscriptsubscript𝑟1𝑚1superscriptsubscript𝛿𝑟2subscript𝑆𝑟superscriptsubscriptproduct𝑗1𝑟superscriptsubscriptℓ𝑟2superscriptsubscript𝑖𝑟𝑚1superscriptsubscriptproduct𝑗𝑖1...
1,580
Multi-Level Composite Stochastic Optimization via Nested Variance Reduction 4 Multi-level nested SPIDER Proof. =6​m2​∑r=1m−1δr2Sr​∏j=1rℓr2​∑i=rm−1(∏j=i+1mℓj2⋅Li+12⋅∏j=1iℓj4)+6​m2​∑r=1m−1τ​ϵ2sr​LF2​∑i=rm−1(∏j=1iℓj4⋅Li+12⋅∏j=i+1mℓj2)absent6superscript𝑚2superscriptsubscript𝑟1𝑚1superscriptsubscript𝛿𝑟2subscript𝑆𝑟supe...
1,601
45
Consider problem (1) with m≥2𝑚2m\geq 2, and suppose Assumptions 2.1 and 4.1 hold. In Algorithm 3, if we set ϵk=m​θ​LFk⋅ℓFsubscriptitalic-ϵ𝑘𝑚𝜃subscript𝐿𝐹⋅𝑘subscriptℓ𝐹\epsilon_{k}=\frac{m\theta L_{F}}{k\cdot\ell_{F}} with θ>0𝜃0\theta>0 and set the other parameters as in Lemma 4.3, then the output x¯¯𝑥\bar{x} sa...
457
Multi-Level Composite Stochastic Optimization via Nested Variance Reduction 4 Multi-level nested SPIDER Theorem 4.4. Consider problem (1) with m≥2𝑚2m\geq 2, and suppose Assumptions 2.1 and 4.1 hold. In Algorithm 3, if we set ϵk=m​θ​LFk⋅ℓFsubscriptitalic-ϵ𝑘𝑚𝜃subscript𝐿𝐹⋅𝑘subscriptℓ𝐹\epsilon_{k}=\frac{m\theta L_{...
483
46
From Theorem 2.3 and Lemma 4.3, the output of Algorithm 3 satisfies , 1 = 𝐄​[‖𝒢​(x¯)‖]≤4​LF​(Φ​(x0)−Φ∗)∑t=0T−1ϵχ​(t)+4​∑t=0T−1ϵχ​(t)2∑t=0T−1ϵχ​(t).𝐄delimited-[]norm𝒢¯𝑥4subscript𝐿𝐹Φsuperscript𝑥0subscriptΦsuperscriptsubscript𝑡0𝑇1subscriptitalic-ϵ𝜒𝑡4superscriptsubscript𝑡0𝑇1subscriptsuperscriptitalic-ϵ2𝜒𝑡su...
1,477
Multi-Level Composite Stochastic Optimization via Nested Variance Reduction 4 Multi-level nested SPIDER Proof. From Theorem 2.3 and Lemma 4.3, the output of Algorithm 3 satisfies , 1 = 𝐄​[‖𝒢​(x¯)‖]≤4​LF​(Φ​(x0)−Φ∗)∑t=0T−1ϵχ​(t)+4​∑t=0T−1ϵχ​(t)2∑t=0T−1ϵχ​(t).𝐄delimited-[]norm𝒢¯𝑥4subscript𝐿𝐹Φsuperscript𝑥0subscrip...
1,498
47
, 1 = ∑k=1K∑i=1m(Bik+(τk−1)​bik)+∑k=1K∑i=1m−1(Sik+(τk−1)​sik)superscriptsubscript𝑘1𝐾superscriptsubscript𝑖1𝑚superscriptsubscript𝐵𝑖𝑘subscript𝜏𝑘1superscriptsubscript𝑏𝑖𝑘superscriptsubscript𝑘1𝐾superscriptsubscript𝑖1𝑚1superscriptsubscript𝑆𝑖𝑘subscript𝜏𝑘1superscriptsubscript𝑠𝑖𝑘\displaystyle\sum_{k=1}^{K...
686
Multi-Level Composite Stochastic Optimization via Nested Variance Reduction 4 Multi-level nested SPIDER Proof. , 1 = ∑k=1K∑i=1m(Bik+(τk−1)​bik)+∑k=1K∑i=1m−1(Sik+(τk−1)​sik)superscriptsubscript𝑘1𝐾superscriptsubscript𝑖1𝑚superscriptsubscript𝐵𝑖𝑘subscript𝜏𝑘1superscriptsubscript𝑏𝑖𝑘superscriptsubscript𝑘1𝐾supersc...
707
48
Consider problem (2) with m≥2𝑚2m\geq 2, and suppose Assumptions 2.1 and 4.1.(a) hold. In addition, let Nmax=max⁡{N1,…,Nm}subscript𝑁maxsubscript𝑁1…subscript𝑁𝑚N_{\mathrm{max}}=\max\{N_{1},\ldots,N_{m}\} and assume the target precision ϵitalic-ϵ\epsilon satisfies Nmax≤ℓF2​m​ϵ.subscript𝑁maxsubscriptℓ𝐹2𝑚italic-ϵ\sqr...
1,059
Multi-Level Composite Stochastic Optimization via Nested Variance Reduction 4 Multi-level nested SPIDER Theorem 4.5. Consider problem (2) with m≥2𝑚2m\geq 2, and suppose Assumptions 2.1 and 4.1.(a) hold. In addition, let Nmax=max⁡{N1,…,Nm}subscript𝑁maxsubscript𝑁1…subscript𝑁𝑚N_{\mathrm{max}}=\max\{N_{1},\ldots,N_{m}...
1,085
49
Through a similar line of proof of Lemma 4.3, 𝐄​[‖vt−F′​(xt)‖2]≤ϵχ​(t)2𝐄delimited-[]superscriptnormsuperscript𝑣𝑡superscript𝐹′superscript𝑥𝑡2superscriptsubscriptitalic-ϵ𝜒𝑡2\mathbf{E}\big{[}\|v^{t}-F^{\prime}(x^{t})\|^{2}\big{]}\leq\epsilon_{\chi(t)}^{2} still holds. Consequently (48) also holds in this case. Not...
1,710
Multi-Level Composite Stochastic Optimization via Nested Variance Reduction 4 Multi-level nested SPIDER Proof. Through a similar line of proof of Lemma 4.3, 𝐄​[‖vt−F′​(xt)‖2]≤ϵχ​(t)2𝐄delimited-[]superscriptnormsuperscript𝑣𝑡superscript𝐹′superscript𝑥𝑡2superscriptsubscriptitalic-ϵ𝜒𝑡2\mathbf{E}\big{[}\|v^{t}-F^{\p...
1,731
50
In Theorem 4.4, if we set ϵk≡ϵsubscriptitalic-ϵ𝑘italic-ϵ\epsilon_{k}\equiv\epsilon, then the Algorithm output x¯¯𝑥\bar{x} s.t. 𝐄​[‖𝒢​(x¯)‖]≤ϵ𝐄delimited-[]norm𝒢¯𝑥italic-ϵ\mathbf{E}[\|\mathcal{G}(\bar{x})\|]\leq\epsilon after K=O​(m​LFϵ⋅ℓF)𝐾𝑂𝑚subscript𝐿𝐹⋅italic-ϵsubscriptℓ𝐹K=O(\frac{mL_{F}}{\epsilon\cdot\ell...
778
Multi-Level Composite Stochastic Optimization via Nested Variance Reduction 4 Multi-level nested SPIDER Remark 4.6. In Theorem 4.4, if we set ϵk≡ϵsubscriptitalic-ϵ𝑘italic-ϵ\epsilon_{k}\equiv\epsilon, then the Algorithm output x¯¯𝑥\bar{x} s.t. 𝐄​[‖𝒢​(x¯)‖]≤ϵ𝐄delimited-[]norm𝒢¯𝑥italic-ϵ\mathbf{E}[\|\mathcal{G}(\ba...
803
51
In this section, we present numerical experiments to demonstrate the effectiveness of the proposed algorithms and compare them with related work. We first apply NPAG with different variance-reduced estimators on a noncovex sparse classification problem. Since this is a one-level finite-sum problem, we also compare it w...
102
Multi-Level Composite Stochastic Optimization via Nested Variance Reduction 5 Numerical experiments In this section, we present numerical experiments to demonstrate the effectiveness of the proposed algorithms and compare them with related work. We first apply NPAG with different variance-reduced estimators on a noncov...
119
52
We consider the following ℓ1subscriptℓ1\ell_{1}-regularized empirical risk minimization problem: , 1 = minw⁡1N​∑i=1Nℓ​(aiT​x,bi)+β​‖x‖1,subscript𝑤1𝑁superscriptsubscript𝑖1𝑁ℓsuperscriptsubscript𝑎𝑖𝑇𝑥subscript𝑏𝑖𝛽subscriptnorm𝑥1\min_{w}\frac{1}{N}\sum_{i=1}^{N}\ell(a_{i}^{T}x,b_{i})+\beta\|x\|_{1},. , 2 = . , 3 ...
1,743
Multi-Level Composite Stochastic Optimization via Nested Variance Reduction 5 Numerical experiments 5.1 Sparse binary classification We consider the following ℓ1subscriptℓ1\ell_{1}-regularized empirical risk minimization problem: , 1 = minw⁡1N​∑i=1Nℓ​(aiT​x,bi)+β​‖x‖1,subscript𝑤1𝑁superscriptsubscript𝑖1𝑁ℓsuperscript...
1,767
53
In this section, we present the numerical results for a risk-averse portfolio optimization problems, which is a common benchmark example used in many stochastic composite optimization methds (e.g.formulae-sequence𝑒𝑔e.g., [16, 41, 42, 21]). Consider the scenario where d𝑑d assets can be invested during the time period...
1,601
Multi-Level Composite Stochastic Optimization via Nested Variance Reduction 5 Numerical experiments 5.2 Sparse portfolio selection problem In this section, we present the numerical results for a risk-averse portfolio optimization problems, which is a common benchmark example used in many stochastic composite optimizati...
1,626
54
For these experiments, CIVR and Nested-SPIDER both take the batch size of ⌈N1/2⌉superscript𝑁12\lceil N^{1/2}\rceil; CIVR-adp takes the adaptive batch size of Sk=⌈min⁡{10​k+1,N1/2}⌉subscript𝑆𝑘10𝑘1superscript𝑁12S_{k}=\lceil\min\{10k+1,N^{1/2}\}\rceil; VRSC-PG and C-SAGA both take the batch size of ⌈N2/3⌉superscript�...
715
Multi-Level Composite Stochastic Optimization via Nested Variance Reduction 5 Numerical experiments 5.2 Sparse portfolio selection problem For these experiments, CIVR and Nested-SPIDER both take the batch size of ⌈N1/2⌉superscript𝑁12\lceil N^{1/2}\rceil; CIVR-adp takes the adaptive batch size of Sk=⌈min⁡{10​k+1,N1/2}⌉...
740
55
We have proposed a normalized proximal approximate gradient (NPAG) method for solving multi-level composite stochastic optimization problems. The approximate gradients at each iteration are obtained via nested variance reduction using the SARAH/Spider estimator. In order to find an approximate stationary point where th...
418
Multi-Level Composite Stochastic Optimization via Nested Variance Reduction 6 Conclusion We have proposed a normalized proximal approximate gradient (NPAG) method for solving multi-level composite stochastic optimization problems. The approximate gradients at each iteration are obtained via nested variance reduction us...
433
56
Under Assumption 4.1.(a), it is straightforward to show that fi=𝐄ξi​fi,ξisubscript𝑓𝑖subscript𝐄subscript𝜉𝑖subscript𝑓𝑖subscript𝜉𝑖f_{i}=\mathbf{E}_{\xi_{i}}f_{i,\xi_{i}} is ℓisubscriptℓ𝑖\ell_{i}-Lipschitz and its gradient fi′=𝐄ξi​fi,ξi′subscriptsuperscript𝑓′𝑖subscript𝐄subscript𝜉𝑖subscriptsuperscript𝑓′𝑖s...
1,846
Multi-Level Composite Stochastic Optimization via Nested Variance Reduction Appendix A Proof of Lemma 4.2 Proof. Under Assumption 4.1.(a), it is straightforward to show that fi=𝐄ξi​fi,ξisubscript𝑓𝑖subscript𝐄subscript𝜉𝑖subscript𝑓𝑖subscript𝜉𝑖f_{i}=\mathbf{E}_{\xi_{i}}f_{i,\xi_{i}} is ℓisubscriptℓ𝑖\ell_{i}-Lips...
1,871
57
, 1 = ‖Fi​(x)−Fi​(y)‖normsubscript𝐹𝑖𝑥subscript𝐹𝑖𝑦\displaystyle\|F_{i}(x)-F_{i}(y)\|. , 2 = =\displaystyle=. , 3 = ‖fi​(Fi−1​(x))−fi​(Fi−1​(y))‖normsubscript𝑓𝑖subscript𝐹𝑖1𝑥subscript𝑓𝑖subscript𝐹𝑖1𝑦\displaystyle\|f_{i}(F_{i-1}(x))-f_{i}(F_{i-1}(y))\|. , 4 = . , 1 = . , 2 = ≤\displaystyle\leq. , 3 = ℓi​‖Fi−...
826
Multi-Level Composite Stochastic Optimization via Nested Variance Reduction Appendix A Proof of Lemma 4.2 Proof. , 1 = ‖Fi​(x)−Fi​(y)‖normsubscript𝐹𝑖𝑥subscript𝐹𝑖𝑦\displaystyle\|F_{i}(x)-F_{i}(y)\|. , 2 = =\displaystyle=. , 3 = ‖fi​(Fi−1​(x))−fi​(Fi−1​(y))‖normsubscript𝑓𝑖subscript𝐹𝑖1𝑥subscript𝑓𝑖subscript𝐹�...
851
58
Algorithm 1 Normalized Proximal Approximate Gradient (NPAG) Method Algorithm 2 Prox-Spider Method Algorithm 3 Multi-level Nested-Spider Method Figure 1: Experiments on sparse binary classification on mnist dataset, β=1/12691𝛽112691\beta=1/12691. Figure 2: Experiments on sparse binary classification on rcv1 dataset, β=...
143
Multi-Level Composite Stochastic Optimization via Nested Variance Reduction References Algorithm 1 Normalized Proximal Approximate Gradient (NPAG) Method Algorithm 2 Prox-Spider Method Algorithm 3 Multi-level Nested-Spider Method Figure 1: Experiments on sparse binary classification on mnist dataset, β=1/12691𝛽112691\...
157
0
This paper augments the reward received by a reinforcement learning agent with potential functions in order to help the agent learn (possibly stochastic) optimal policies. We show that a potential-based reward shaping scheme is able to preserve optimality of stochastic policies, and demonstrate that the ability of an a...
170
Potential-Based Advice for Stochastic Policy Learning Abstract This paper augments the reward received by a reinforcement learning agent with potential functions in order to help the agent learn (possibly stochastic) optimal policies. We show that a potential-based reward shaping scheme is able to preserve optimality o...
181
1
Reinforcement learning (RL) is a framework that allows an agent to complete tasks in an environment, even when a model of the environment is not known. The agent ‘learns’ to complete a task by maximizing its expected long-term reward, where the reward signal is supplied by the environment. RL algorithms have been succe...
721
Potential-Based Advice for Stochastic Policy Learning I Introduction Reinforcement learning (RL) is a framework that allows an agent to complete tasks in an environment, even when a model of the environment is not known. The agent ‘learns’ to complete a task by maximizing its expected long-term reward, where the reward...
733
2
Shaping or augmenting the reward received by an RL agent in order to enable it to learn optimal policies faster is an active area of research. Reward modification via human feedback was used in [9, 10] to interactively shape an agent’s response so that it learned a desired behavior. However, frequent human supervision ...
449
Potential-Based Advice for Stochastic Policy Learning II Related Work Shaping or augmenting the reward received by an RL agent in order to enable it to learn optimal policies faster is an active area of research. Reward modification via human feedback was used in [9, 10] to interactively shape an agent’s response so th...
462
3
An MDP [27] is a tuple (S,A,𝕋,ρ0,R)𝑆𝐴𝕋subscript𝜌0𝑅(S,A,\mathbb{T},\rho_{0},R). S𝑆S is the set of states, A𝐴A the set of actions, 𝕋:S×A×S→[0,1]:𝕋→𝑆𝐴𝑆01\mathbb{T}:S\times A\times S\rightarrow[0,1] encodes ℙ​(st+1|st,at)ℙconditionalsubscript𝑠𝑡1subscript𝑠𝑡subscript𝑎𝑡\mathbb{P}(s_{t+1}|s_{t},a_{t}), the p...
1,126
Potential-Based Advice for Stochastic Policy Learning III Preliminaries III-A Reinforcement Learning An MDP [27] is a tuple (S,A,𝕋,ρ0,R)𝑆𝐴𝕋subscript𝜌0𝑅(S,A,\mathbb{T},\rho_{0},R). S𝑆S is the set of states, A𝐴A the set of actions, 𝕋:S×A×S→[0,1]:𝕋→𝑆𝐴𝑆01\mathbb{T}:S\times A\times S\rightarrow[0,1] encodes ℙ​(...
1,146